Invented by Alexander Krowitz, Martin Tapp, Kronos Technology Systems Ltd Partnershiip, Kronos Technology Systems LP
The Kronos Technology Systems Ltd Partnershiip, Kronos Technology Systems LP invention works as followsA computer-implemented training method for machine learning models is presented. The method involves collecting historical data from a database and applying one or multiple transformations on the historical data in order to create a model feature set. It then separates the model feature set into one or several pools, with each pool containing one or many model features that are homogeneous based on a common value. The method also includes creating dynamically a training data set for each pool that includes one or more model features from the pool as well as at least some historical data. For each training set, the method includes training a machine-learning model on that training set.
Background for Method for training machine-learning models to make simulated estimates
The invention relates to the use of machine learning for accurate forecasting.
This section provides a context or background. The description can include concepts that have been conceived, but not necessarily pursued. “Unless otherwise indicated, what is described here is not considered prior art for the description and claims. It is also not admitted as prior art through inclusion in this part.
Machine learning regression can be used to model numerical patterns using historical data. Machine learning regression can be based on ‘training? Machine learning regression may use?training? Beispiele to capture characteristics of their unknown probability distribution. Regression methods can dynamically create complex formulas for predicting business patterns. The training data can be viewed as examples of relationships between variables. A ‘pool’ of data can also be used to train, as machine learning algorithms can determine feature combinations dynamically. In the current environment, data can be gathered from many stores or departments that are similar to the one being predicted. The systems can detect patterns which may be uncommon in a particular store but are common in the entire organization and use the pattern detected in future predictions.
Some conventional systems use business volume forecasts in order to determine the workload required, but these systems don’t integrate multiple data sources.
Other conventional systems focus on atypical event and/or use neural net architecture for external features.
Businesses that use these systems spend resources manually correcting schedules because of the inaccuracy.
What is needed is to find a way to use machine-learning to produce accurate forecasts, without the inherent problems of the previous systems. These improved forecasts can reduce such expenditures and allow employees and organizations to focus on their core missions.
The summary below is only representative and not limiting.
The embodiments can be used to solve the above problems and realize other benefits.
While prior approaches might rely on fixed inputs and static formulas, machine learning methods can be dynamic. In some situations, features such as windowed trends or seasonality are irrelevant for predicting business volumes. They can be ignored and in other situations they are useful. A machine learning method is able to use significantly more data and features than static formulas. The training process also benefits from the expansion of data, as it combines data from multiple stores or departments to create a single “pool”. This complex model is used to predict for each unit in the prediction phase. It is also possible to easily incorporate new types of historical information into the modeling process. Data from third parties, like weather data and local events calendars, that are not part of the core business can be easily added as they become available.
In a second aspect, a method is provided that uses machine-learning regression to predict retail business volume using historical data. The method can be used to predict different types of retail business volume including sales volume and transaction volume. This method has two stages: training and prediction. In both stages, historical data is transformed into model features using multiple transformations. Data from the past may include business volume and other data types. The data may include store characteristics such as department, geographical location, weather, climate and local data. Model features can represent the exact values or transformations, such as those that capture seasonality, trends and effects of special events, like sales or store closings.
The training phase uses machine learning regression to create models that embody meaningful patterns extracted from historic data. The prediction phase uses the model on the most recent data in order to make volume predictions. Previous predictions can be used as a “backfill” if complete historical data are not available. The historical data can be used to make predictions based on the current data. The system could also have a monitoring component that can identify where system performance is lacking.
Improved forecasting can lead to better staffing decisions. Retail businesses, for example, may improve their customer experience and the efficiency and effectiveness of their operations and transaction. Correct scheduling leads to better resource use for employers, and thus reduced costs. This also leads to a better employee experience and therefore better retention.
BRIEF DESCRIPTION ABOUT THE VIEWS FROM THE DRAWINGS
When read with the accompanying figures, the description will make more sense.
FIG. “FIG.Click here to view the patent on Google Patents.